A Linear Predictive Front-End Processor for Speech Recognition in Noisy Environments

We describe a recently developed algorithm for autoregressive (AR) modeling of speech signals which have been degraded by uncorrelated additive noise [1], and examine the algorithm performance in a speech recognition system which uses linear predictive analysis. The proposed algorithm aims at modeling the original source through a composite modeling of the noisy source by an AR model for the clean source, and an additive, appropriately chosen, model for the noise.